What is the best LLM for a real-time voice agent?

The board ranks each parameter on its own and publishes no single winner, and on this run the parameters disagree. OpenAI gpt-5.6-luna each completed 94.21% of the multi-step tool tasks while staying silent on most of their tool calls, whereas Anthropic Claude Haiku 4.5 completed 82.46% and goes silent on only 2.7%.

Which LLM has the lowest first-token latency?

Baseten inkling-small measured 177ms TTFT p50 from us-east4, the lowest on the board. Cerebras gpt-oss-120b measured 195ms but stays silent on 100.0% of tool calls.

How fast does an LLM need to respond for voice?

Every millisecond before the first token is silence to the caller, so the board measures first-token latency as its own parameter. Measured p50s range from 177ms (Baseten inkling-small) to 2.15s (xAI Grok-4.3), which is too slow to lead a live turn.

Which LLMs fabricate answers in a voice agent?

Fabrication is inventing a policy or fact the agent was never given. OpenAI gpt-4.1-mini measured 57% on out-of-policy questions and Alibaba qwen-turbo 87%, while Anthropic Claude Haiku 4.5 measured 0%.

What is dead air in a voice agent?

A turn where the caller hears silence instead of speech. Anthropic Claude Haiku 4.5 measured 1.6% of turns against 68.2% for Cerebras gpt-oss-120b, but a quiet agent is not the same as a working one: Alibaba qwen-turbo speaks on almost every turn and still finished only 6.58% of the tool tasks, because it narrates instead of acting.

LLM · Speko Benchmarks

Task done

higher is better

Did the agent finish the job without doing damage — look the order up, refund the right amount, and leave alone what it must not touch?

94.21%

92.46%

91.40%

89.12%

87.72%

87.37%

87.19%

86.32%

85.00%

82.46%

81.72%

80.18%

77.72%

77.37%

74.74%

72.54%

72.46%

72.02%

71.14%

64.21%

57.72%

32.37%

6.58%

0%25%50%75%100%

Stalled

lower is better

The agent says it will do something, then goes quiet without doing it. The caller has to ask again.

Grok-4.37%

gpt-5.6-luna8%

gpt-oss-120b9%

gpt-4.114%

Qwen3.8-27B14%

gpt-5-mini21%

gpt-526%

DeepSeek-V4-Flash-073128%

gemini-3.5-flash-lite28%

gemini-3.8-flash29%

gemma-4-31b39%

inkling42%

GLM-4.743%

inkling-small46%

Nemotron-3-Ultra54%

gpt-5-nano87%

qwen-turbo94%

0%25%50%75%100%

Fabrication

lower is better

The agent states a fact it was never given — an opening time, a delivery window, a policy. It sounds confident and it is invented.

11 models0 of 30≤31% · 10 independent0%

Claude Haiku 4.5

Claude Sonnet 5

gemini-3.8-flash

gemma-4-31b gpt-5-mini gpt-5.6-luna gpt-5.6-terra Grok-4.3

inkling-small

Llama-3.3-70B

Qwen3.8-27B

qwen-turbo26 of 3087%

gpt-4.1-mini17 of 3057%

gpt-oss-120b9 of 3030%

Nemotron-3-Ultra8 of 3027%

gpt-oss-120b8 of 3027%

GLM-4.75 of 3017%

gpt-5-nano5 of 3017%

DeepSeek-V4-Flash-07313 of 3010%

gemini-3.5-flash-lite3 of 3010%

gpt-4.13 of 3010%

gpt-53 of 3010%

inkling1 of 303%

Refusal quality

higher is better

When the agent does not know something, does it say what it is missing and offer a way forward, or just stop? Scored on the same probes as Fabrication.

Gemini3.8-flash96.7%

xAIGrok-4.396.7%

OpenAIgpt-5.6-luna95.6%

AnthropicClaude Sonnet 590.0%

Cerebrasgemma-4-31b90.0%

OpenAIgpt-5.6-terra88.9%

AnthropicClaude Haiku 4.584.4%

OpenAIgpt-584.4%

OpenAIgpt-5-nano84.4%

OpenAIgpt-4.1-mini83.3%

OpenAIgpt-5-mini83.3%

Cerebrasgpt-oss-120b83.3%

BasetenGLM-4.780.0%

Basetengpt-oss-120b80.0%

CerebrasQwen3.8-27B80.0%

OpenAIgpt-4.178.9%

Baseteninkling-small77.8%

BasetenNemotron-3-Ultra75.6%

Baseteninkling74.4%

BasetenDeepSeek-V4-Flash-073172.2%

Gemini3.5-flash-lite70.0%

TogetherLlama-3.3-70B20.0%

Alibabaqwen-turbo11.1%

0%25%50%75%100%

Silence

lower is better

How often the caller hears nothing at all — every turn counted, not just the ones where a tool runs. The small number is the narrower measure: the share of tool calls the agent made in complete silence, which for 10 of the 22 measured models is all of them.

Claude Haiku 4.52%3% of tool calls4 of 248 turns

GLM-4.72%4% of tool calls6 of 260 turns

DeepSeek-V4-Flash-07315%7% of tool calls16 of 314 turns

qwen-turbo9%100% of tool calls ≥16%10 of 110 turns

Qwen3.8-27B33%49% of tool calls98 of 295 turns

gpt-5-nano39%100% of tool calls ≥80%65 of 165 turns

Claude Sonnet 540%59% of tool calls117 of 294 turns

gpt-4.1-mini40%66% of tool calls102 of 255 turns

gemma-4-31b53%95% of tool calls122 of 229 turns

Grok-4.355%83% of tool calls155 of 283 turns

Nemotron-3-Ultra56%94% of tool calls139 of 248 turns

gpt-4.159%94% of tool calls159 of 270 turns

inkling66%94% of tool calls203 of 306 turns

gpt-5.6-terra66%100% of tool calls ≥82%195 of 294 turns

gpt-5-mini67%100% of tool calls ≥82%198 of 295 turns

gpt-oss-120b68%100% of tool calls ≥82%204 of 299 turns

gemini-3.5-flash-lite69%100% of tool calls ≥82%197 of 287 turns

gemini-3.8-flash69%100% of tool calls ≥82%214 of 310 turns

gpt-5.6-luna70%100% of tool calls ≥82%230 of 330 turns

gpt-oss-120b71%100% of tool calls ≥82%227 of 319 turns

gpt-572%100% of tool calls ≥82%216 of 301 turns

inkling-small72%95% of tool calls244 of 339 turns

Llama-3.3-70B74%100% of tool calls ≥82%260 of 350 turns

0%25%50%75%100%

Tool-arg fabrication

lower is better

The prompt orders a lookup the agent has no id for. It calls the tool anyway and fills the argument in — blank, a placeholder, or a number it made up.

6 models0 of 300%

Claude Sonnet 5

DeepSeek-V4-Pro gpt-4.1 Grok-4.3